An in-depth look at Google’s newest (and oldest) cloud

Compute Engine is the newest kid on Google’s block of cloud products. It sits in the same suite as App Engine, Big Query, and other products that allow Google to make money on what it does best: big, hairy server/data center work.

We just sat down for a lovely chat with Urs Hölzle (pictured), Google’s senior vice president of infrastructure, who presented Compute Engine to a thrilled audience of developers today at Google I/O.

First things first, we asked him how Compute Engine fits into Google’s other cloud offerings.

“What’s the difference between a platform and an infrastructure? Nobody knows,” Hölzle quipped, quickly adding that while App Engine and Big Query allow developers a lot of convenience, Compute Engine allows them to get closer to the metal, meaning they can see better performance for large-scale computational tasks.

For example, in his onstage demo, Hölzle showed Compute Engine crunching human genome data to find associations between chromosomes. It took Compute Engine seconds or fractions of a second to perform tasks that used to take 10 minutes. “That’s what happens when you have 10,000 cores working on it…. There are 771,886 cores available to the app right now,” he said during the demo.

“That is how infrastructure as a service is supposed to work.”

In our offstage talk, he said of that particular party trick, “No one’s demonstrated anything even remotely close to what we showed in our live demo today. That speaks for the quality of the underlying infrastructure…. Performance was the first thing our beta customers commented on.”

He continued to say that at least some of that performance is due to the fact that the developers are much closer to the hardware. Part of Compute Engine is that the developer has the duty or ability, depending on your perspective, of certain administrative tasks not within the grasp of an App Engine customer.

“These other things [Google’s other cloud services] are more like services,” Hölzle said. “We do a lot of the [sysadmin] work for you…. It’s not like one or the other is better, they’re just different tools.”

Compute Engine, while it is intended for use by customers with some experience in cloud computing, is well suited to any number of tasks. “Infrastructure, because it is so low level, there’s really no intrinsic bias toward one workload or another,” Hölzle said.

Right now, Compute Engine is being used for big-data crunching, large-scale computations, like video transcoding, but Hölzle said the service is valuable for anyone needing a high level of performance, scale, and consistency.

He also said Compute Engine’s cost savings were a draw. “We have seen beta customers that spent up to 50 percent less than [with] their equivalent setup with another cloud provider,” he said.

These are big claims to make, and claims that any cloud services provider would make: We’re big. We’re stable. We’re cheap.

But Hölzle, who was one of the search giant’s first ten employees, backs up Google’s particular claims with an interesting anecdote.

“Search is such a computationally intensive job,” he said. “From the beginning, the biggest problem in the first two years at Google was scale…. When I started, [Google web search] was university code. A search at noon on a Monday would not work.”

So his first task at Google was really figuring out how to build an infrastructure that would allow Google’s nascent web search service to scale.

“We also had to deal with cost,” he said. “Our monetization back then was not high, and a lot of our products are free. If you don’t make sure that the infrastructure costs a fraction of the low revenue per user, you’re in trouble.

“It’s natural, it’s easy for us to do well for an external offering because we are familiar with these topics from the last century…. This is the externalization of what we’ve been working on for a long time.”

Google Compute Engine is available now for a limited audience. Currently, Google is seeking customers that would need to use 100 or more virtual machines. General availability will roll out, well, later, “depending on how well it’s going,” Hölzle said.